Anthropic explains its planned text watermark for Claude

Anthropic said on August 14 that future Claude models will generate text carrying a statistical watermark. Its technical explanation says the system uses a version of Google DeepMind’s SynthID-Text approach, changing the randomness used when several next-token choices are similarly suitable. It does not insert hidden characters or identifying account data.

Detection is probabilistic and content-dependent

With the secret detection key, the method estimates the likelihood that Claude participated in producing a passage. Anthropic explicitly says this is not proof of human authorship, exclusive Claude authorship, or the absence of another model. Confidence improves with longer passages because more eligible choices contribute signal.

Factual prose, light proofreading, and code offer fewer interchangeable token choices, so they can carry less detectable signal. Rewriting can also weaken or remove it. Those limitations matter for product requirements: a detector result should not become the sole trigger for discipline, takedown, fraud labeling, or authorship decisions.

Integrations need provenance as well as detection

Anthropic says it plans to apply the watermark globally and was releasing a detection API in private preview to eligible organizations. For supported image and document types, the company describes a different mechanism: signed C2PA content credentials in file metadata.

Developers should separate three signals in their data model: provider-signed provenance, probabilistic text detection, and user-declared authorship. Store the detector version, threshold, passage length, and confidence rather than a boolean label. Test transformations such as formatting, translation, code extraction, and copy editing against the intended workflow. Where consequences are significant, require corroborating evidence and human review; the source’s own limitations make that an engineering necessity.